No description available.

Anomaly detection aims to identify the data samples that do not conform to a known normal (regular) behavior. As the definition of an anomaly is often ambiguous, unsupervised and semi-supervised deep learning (DL) algorithms that primarily use unlabeled datasets to model normal (regular) behaviors, are popularly studied in this context. The unmanned aerial system (UAS) can use contextual anomaly detection algorithms to identify interesting objects of concern in applications like search and rescue, disaster management, public security etc. This thesis presents a novel multi-stage framework that supports detection of frames with unknown anomalies, localization of anomalies in the detected frames, and validation of detected frames for incremental semi-supervised learning, with the help of a human operator. The proposed architecture is tested on two new datasets collected for a UAV-based system. In order to detect and localize anomalies, it is important to both model the normal data distribution accurately as well as formulate powerful discriminant (anomaly scoring) techniques. We implement a generative adversarial network (GAN)-based anomaly detection architecture to study the effect of loss terms and regularization on the modeling of normal (regular) data and arrive at the most effective anomaly scoring method for the given application. Following this, we use incremental semi-supervised learning techniques that utilize a small set of labeled data (obtained through validation from a human operator), with large unlabeled datasets to improve the knowledge-base of the anomaly detection system. / Master of Science / Anomaly detection aims to identify the data samples that do not conform to a known normal (regular) behavior. As the definition of an anomaly is often ambiguous, most techniques use unlabeled datasets, to model normal (regular) behaviors. The availability of large unlabeled datasets combined with novel applications in various domains, has led to an increasing interest in the study of anomaly detection. In particular, the unmanned aerial system (UAS) can use contextual anomaly detection algorithms to identify interesting objects of concern in applications like search and rescue (SAR), disaster management, public security etc. This thesis presents a novel multi-stage framework that supports detection and localization of unknown anomalies, as well as the validation of detected anomalies, for incremental learning, with the help of a human operator. The proposed architecture is tested on two new datasets collected for a UAV-based system. In order to detect and localize anomalies, it is important to both model the normal data distribution accurately and formulate powerful discriminant (anomaly scoring) techniques. To this end, we study the state-of-the-art generative adversarial networks (GAN)-based anomaly detection algorithms for modeling of normal (regular) behavior and formulate effective anomaly detection scores. We also propose techniques to incrementally learn the new normal data as well as anomalies, using the validation provided by a human operator. This framework is introduced with the aim to support temporally critical applications that involve human search and rescue, particularly in disaster management.

As the number of surveillance cameras and mobile platforms with cameras increases, automated detection and tracking of objects on these systems gain importance. There are various tracking methods designed for stationary or moving cameras. For stationary cameras, correspondence based tracking methods along with background subtraction have various advantages such as enabling detection of object entry and exit in a scene. They also provide robust tracking when the camera is static. However, they fail when the camera is moving. Conversely, histogram based methods such as mean shift enables object tracking on moving camera cases. Though, with mean shift object&rsquo / s entry and exit cannot be detected automatically which means a new object&rsquo / s manual initialization is required. In this thesis, we propose a dual-mode object tracking system which combines the benefits of correspondence based tracking and mean shift tracking. For each frame, a reliability measure based on background update rate is calculated. Interquartile Range is used for finding outliers on this measure and camera movement is detected. If the camera is stationary, correspondence based tracking is used and when camera is moving, the system switches to the mean shift tracking mode until the reliability of correspondence based tracking is sufficient according to the reliability measure. The results demonstrate that, in stationary camera mode, new objects can be detected automatically by correspondence based tracking along with background subtraction. When the camera starts to move, generation of false objects by correspondence based tracking is prevented by switching to mean shift tracking mode and handing over the correct bounding boxes with a seamless operation which enables continuous tracking.

La détection d'objets mobiles dans des flux vidéo est une étape essentielle pour de nombreux algorithmes de vision par ordinateur. Cette tâche se complexifie lorsque la caméra utilisée est en mouvement. En effet, l'environnement capté par ce type de caméra apparaît en mouvement et il devient plus difficile de distinguer les objets qui effectuent réellement un mouvement de ceux qui constituent la partie statique de la scène. Dans cette thèse, nous apportons des contributions au problème de détection d'objets mobiles dans le flux vidéo d'une caméra mobile. L'idée principale qui nous permet de distinguer les éléments mobiles de ceux qui sont statiques repose sur un calcul de distance dans l'espace 3D. Les positions 3D de caractéristiques extraites des images sont estimées par triangulation puis leurs mouvements 3D sont analysés pour réaliser un étiquetage éparse statique/mobile de ces points. Afin de rendre la détection robuste au bruit, l'analyse des mouvements 3D des caractéristiques est comparée à d'autres points précédemment estimés statiques. Une mesure de confiance, mise à jour au cours du temps, est utilisée pour déterminer l'étiquette à attribuer à chacun des points. Nos contributions ont été appliquées à des jeux de données virtuelles (issus du projet Previz 2) et réelles (reconnus dans la communauté [Och+14]) et les comparaisons ont été réalisées avec l'état de l'art. Les résultats obtenus montrent que la contrainte 3D proposée dans cette thèse, couplée à une analyse statistique et temporelle des mouvements, permet de détecter des éléments mobiles dans le flux vidéo d'une caméra en mouvement et ce même dans des cas complexes où les mouvements apparents de la scène ne sont pas uniformes / Moving objects detection in video streams is a commonly used technique in many computer vision algorithms. The detection becomes more complex when the camera is moving. The environment observed by this type of camera appeared moving and it is more difficult to distinguish the objects which are in movement from the others that composed the static part of the scene. In this thesis we propose contributions for the detection of moving objects in the video stream of a moving camera. The main idea to differenciate between moving and static objects based on 3D distances. 3D positions of feature points extracted from images are estimated by triangulation and then their 3D motions are analyzed in order to provide a sparse static/moving labeling. To provide a more robust detection, the analysis of the 3D motions is compared to those of feature points previously estimated static. A confidance value updated over time is used to decide on labels to attribute to each point.We make experiments on virtual (from the Previz project 1) and real datasets (known by the community [Och+14]) and we compare the results with the state of the art. The results show that our 3D constraint coupled with a statistical and temporal analysis of motions allow to detect moving elements in the video stream of a moving camera even in complex cases where apparent motions of the scene are not similars

Motion segmentation is an important process for separating moving objects from the background. Common methods usually assume fixed camera, other approaches exist as well, but they are usually very computational intensive. This work presents an approach for scene segmentation to regions with coherent motion, which works faster than similar methods and it is capable of online processing with no prior knowledge of objects or camera. The main assumption is that the points belonging to a single objects are moving together and this applies as well in the opposite direction. The proposed method is based on tracking of feature points and searching for groups with similar motion by using RANSAC-based algorithm. Short-range repair of broken tracks is applied to increase the overall robustness of tracking. Found clusters are subsequently processed to represent separate moving objects.